This paper proposes a hybrid anomaly detection framework for Distributed File System (DFS) log analysis, combining statistical machine learning with deep learning sequence models. Rather than relying on a single classifier, the system integrates Random Forest, Linear Support Vector Machine (SVM), Bidirectional Long Short-Term Memory (BiLSTM), and DeepLog within a weighted soft-voting ensemble. Class imbalance is addressed through the Synthetic Minority Over-sampling Technique (SMOTE), applied to classical learners prior to training. Precision-Recall Area Under the Curve (PR-AUC) is adopted as the primary evaluation metric, given its suitability for imbalanced classification tasks. The dual-modality design captures both event-frequency patterns via Bag-of-Words representations and temporal ordering dependencies via sequential models. Experiments on the public HDFS LogHub dataset show the ensemble achieves a PR-AUC of 0.767 and perfect precision, while Random Forest attains the highest F1-score and a ROC-AUC of 0.953. The framework provides a reliable and interpretable approach to log-based anomaly detection in large-scale distributed systems.

Keywords—System log file Analysis; Anomaly Detection; Hybrid Machine Learning; Deep Learning; Log-based Monitoring; Random Forest; BiLSTM(Ensemble); DeepLog(Ensemble); SMOTE (Class Imbalance); Precision-Recall AUC (PR-AUC).